Home › Multilingual speech data
Multilingual speech data, recruited where the speakers actually are
The gap in multilingual datasets is not the recording. It is that nobody went and found the speakers. We recruit directly for the languages, dialects and code-switched speech that corpora keep missing.
Standard forms are not what people speak
Most multilingual datasets are built around the written standard of a language, because that is what is easy to script and easy to check. Users do not speak the written standard. They speak a regional variety, often mixed with another language, at a pace and with a vocabulary the corpus never contained.
So we specify at the level of the spoken variety: Egyptian rather than Modern Standard Arabic, Flemish rather than Netherlands Dutch, Quebec rather than metropolitan French, Brazilian rather than European Portuguese - and code-switched speech collected on purpose rather than filtered out as noise. The languages page sets out where we recruit.
How the projects run
- Native-speaker direction. Sessions are run by someone who speaks the language, because direction is the difference between natural speech and a nervous read.
- Native-speaker QC. Review by speakers of the variety, not just format validation.
- Consistent technical spec across every language. A multilingual set that is technically inconsistent between languages introduces exactly the confound you were trying to avoid.
- Balanced demographics per language. Age and gender balance done per language rather than averaged across the whole set, which is how imbalance hides.
Coverage is a compliance question now
Under the EU AI Act, data governance for high-risk systems includes examining datasets for gaps and possible bias. Language and dialect coverage is one of the most visible places that examination lands, and “we used what was available” is not an answer that survives it. Designing coverage into recruitment is cheaper than explaining its absence later - see Article 10 in practice.
Production, capacity and European language studio detail sit with our parent group at 11c Media.
Multilingual sourcing questions
Which languages can you source?
We recruit across European, South Asian, Southeast and East Asian, Middle Eastern, North African and Sub-Saharan African languages, plus a wide spread of English accents. The full picture is on the languages page - and if a language is not listed, ask, because the list follows demand.
Can you cover dialects rather than standard forms?
That is usually the actual requirement. Modern Standard Arabic does not help a system whose users speak Egyptian or Levantine. Netherlands Dutch does not cover Flemish. We recruit for the spoken variety, not the written standard.
How do you quality-control a language nobody in-house speaks?
With native-speaker direction in session and native-speaker review after it. Quality control on speech data cannot be done by someone who does not speak the language, and anyone claiming otherwise is checking file formats and calling it QC.
Can you handle code-switching?
Yes, and it is increasingly the point. Hinglish, Taglish, Arabic-French and similar mixed speech is how enormous numbers of people talk and is almost absent from commercial corpora. Collecting it deliberately is far more useful than hoping it appears.
Name the languages and the varieties
We will tell you what is recruitable, in what timeframe, and what it costs.